Deep learning-based fake face detection: a comprehensive approach for model development, training optimization, and performance evaluation

This thesis study has been conducted with the aim of developing an effective detection system against the escalating threat of deepfakes. The capability of generative adversarial networks and convolutional neural networks to produce realistic fake facial content poses serious challenges for cybersecurity and social media platforms. In this research, the ConvNeXt architecture has been optimized for deepfake detection, and the model's detection capability has been preserved through an innovative stage configuration with 61% fewer parameters (from 29.1M to 11.3M). The Squeeze-and-Excitation attention mechanism has been employed to enable the network to focus on discriminative regions. The combination of ensemble learning methods and Test-Time Augmentation techniques has provided significant improvements in detection accuracy. The real image dataset consists of frames from FaceForensics++ and CelebA datasets, while manipulated images have been obtained from fake videos (Deepfakes, Face2Face, FaceSwap, NeuralTextures) in the FaceForensics++ dataset. Precision, recall, F1-score, and AUC-ROC metrics have been utilized for performance evaluation. Experimental results have demonstrated that the proposed ensemble-based approaches exhibit superior performance. This study constitutes a significant step in developing effective defense mechanisms against deepfake threats by offering contributions in parameter-efficient model design and ensemble optimization strategies.